I'm starting to practice TDD in my project, as a background it also contains legacy code. We use Mockito as a mocking framework and follow a Spring MVC approach.
There are times when there's a Service class implemented with many different DAO objects as #Autowired properties. There are simple methods within these services, like for example completeTransaction.
completeTransaction will use many of the DAO objects to complete its responsibilities
Update and save the transaction
Advance a business process
Closing other pending operations
However, in performing those operations, the method requires calls to different DAO to fetch and update the transaction, fetch a business process ID, fetch pending transactions (and save their updates). This means that unit testing this method makes me add many #Mock properties. And I need to set up the mock objects before the test will actually finish for me to test a certain condition.
This seems like a code smell, and to me it almost feels like the test is ensuring the implementation of the code instead of just its contract. Again, without mocking the dependencies, the test case will not run (due to NPE and others).
What is a strategy that I can follow to clean up code like this? (I can't really provide the actual source code on the question though). I'm thinking that one possibility would be to set up a facade class with methods like ("getPendingOperations" and "advanceBusinessProcess"). Then I can mock a single dependency. But then I figure that in all other classes that have situations like this I would need to do the same, and then I'm afraid to end up with a lot of "helper" classes just for the sake of cleaner tests.
Thank you in advanced.
I think you'll want to do two things in general when you find yourself with too many mocks. These are not necessary easy, but you may find them helpful.
1) Try and make your methods and classes smaller. I think Clean Code says there are two rules, that classes should small. And that classes should be smaller then that. This makes some sense because as the units you are testing (methods and classes) get smaller, so will the dependencies. You will of course end up with more tests, but they will have less setup in each test.
2) Look at the Law of Demeter (https://en.wikipedia.org/wiki/Law_of_Demeter). There are a bunch of rules, but basically, you want to avoid long string of property/method calls. objA = objB.propertyA.SomeMethod().propertyC; If you need to mock out all of these objects just to get objA, the you will have a lot of setup. But if you can replace this with objA = objB.newProperty; then you only need to mock objB and it's one property.
Neither of these are silver bullets, but hopefully you can use some of these ideas with your project.
If the unit test is testing the completeTransaction method, then you must mock everything on which it depends. Since you are using Mockito, you can use verify to confirm that the correct mocked methods are called and that they are called in the correct order.
If the unit test is testing something that calls the completeTransaction method, then just mock the completeTransaction method.
If this is your class hierarchy:
class A -> class B -> class C
(where -> is "depends on")
In unit tests for class A, mock only class B.
In unit tests for class B, mock only class C.
Related
I am very new to Python world and learning my way in it. I have Java and C# background and in Java/C# unit testing approach is very clear. We have a class under test and we inject all the mocked dependency with expectations to test code.
like:
Car class with a service fuelLevel
Test case: Car is showing low fuel indication, we can write a test.
FuelLevelService fsMock = mock(FuelLevelService .class);
when(fsMock.getLevel()).thenReturn(LOW);
Car car = Car(fsMock)
car.start()
assertTrue(car.warningIndicator());
assertTrue(car.warning().contains("Fuel level low Warning"));
In python i have read we can use monkey patching to set the dependent objects and we can test.
#mock.patch('models.fuel_service.getLevel')
def test_main(self, fuel_service):
fuel_service.return_value = 'LOW'
Car car = Car(fuel_service)
car.start()
self.assertEqual(car.warningIndicator(), True)
I was going to use this approach. But my colleagues have rejected this idea and are proposing to instantiate the actual objects (FuellevelService) and set its value. I find it to be very cumbersome and not needed as in mock also we can do the same with just our expectations.
I am currently struggling to understand if i am wrong or in python my colleagues approach is regarded better pattern.
*Above i have given a simple example. In our project to create an instance object and setting its value is 10-25 lines of code [don't ask me why but that is how it is :)] To create this object and set a value so it returns that same value looks extra maintenance to me. For this approach to create dependent object we will have many py files in module.
Another point is we are here only testing class under test (not dependency service class), class under test which will react on the returned value of the dependency service. We can can have both negative and positive tests to cover class under test.
For dependency service class in same way it should have its own tests to validate all its scenarios separability. So it will be class under test there. By doing this we will have clear separation and test will fail where there is bug.
You only need to use mocks if creating the real objects are inconvenient/cumbersome to create (see here).
If, as you allude to in your question, you can create the real instance cheaply, and set its value easily, there's no real need for a mock.
Additionally, you don't have any guarantee that the mock behaves like the real implementation, so your test might test... something, but not necessarily how your code will really behave in production. Using the production class is preferable.
I'm currently working with mocking with Mockito using jUnit and I've stumbled upon the Partial Mocking section where you use Mockito.spy to partially mock the object. I don't seem to understand this concept of partial mocking since I can't find a scenario why I should use it (since it's pretty similar to mocking in general).
Can anybody explain how partial mocking differs from the normal mocking? And if possible, kindly provide examples.
Thanks!
Partial mocking is where you take a class and ask it to behave as normal, except you want to override certain functionality.
This is useful for unit testing services who communicate with other parts of your application. By overriding the behaviour that would call the other part of your application you can test your service in isolation.
Another example would be when a component would communicate with a database driver. By mocking the part that would communicate with the driver, you can test that part of the application without having to have a database.
From the EasyMock 2.2 classextension documentation:
Sometimes you may need to mock only some methods of a class and keep
the normal behavior of others. This usually happens when you want to
test a method that calls some others in the same class. So you want to
keep the normal behavior of the tested method and mock the others.
I sometimes use this to mock (complicated or process intensive) private methods that are allready fully tested.
Partial mocking can be very handy, but I try to avoid it as much as possible.
Partial mocking:
Say you have a class which takes 10+ parameters for a constructor (this shouldn't ever happen but for this example lets say it does) it's a real chore to create that entire object. Frameworks like mockito let you just use the parts of the object you really want to test.
for example
#Mock BigClass big; //contains loads of attributes
...
when(big.getAttributeOneOfTwenty()).thenReturn(2); //these are static imports from mockito
I find it useful when I'm forced to work with APIs relying on inheritance from abstract classes and/or legacy code working with nonmockable static classes (one real life example - DAO).
Partial mocking (in sense of using the Spy facility from Mockito) allows you to mock calls to inherited methods in the first case, or wrap calls to static methods you are forced to use into normal methods that you can mock, verify etc.
Generally you should design and write code in such a way, that you won't need this (dependency injection, single responsibility per class etc). But from time to time it's useful.
A quick and rough example, to visualize the static API example:
class BigUglyStaticLegacyApi {
public static Foo someStaticMethodFetchingFoo() {...}
}
class Bar {
public void someMethodYouTest() {
Foo foo = getFoo();
//do something with Foo (a FooBar, for example :) )
}
/*this one you mock via spying - not the most elegant solution,
but it's better than nothing */
#VisibleForTesting
protected Foo getFoo() {
return BigUglyStaticLegacyApi.someStaticMethodFetchingFoo();
}
}
I use it the most to mock some methods in my CUT (Class Under Test) but not the method/s I'm actually unit testing. It is an important feature that should be used in unit testing with Mockito.
In following code the issue is, that I cannot test dao.add() without using dao.list().size() and vice versa.
Is this approach normal or incorrect? If incorrect, how can it be improved?
public class ItemDaoTest {
// dao to test
#Autowired private ItemDao dao;
#Test
public void testAdd() {
// issue -> testing ADD but using LIST
int oldSize = dao.list().size();
dao.add(new Item("stuff"));
assertTrue (oldSize < dao.list().size());
}
#Test
public void testFind() {
// issue -> testing FIND but using ADD
Item item = new Item("stuff")
dao.add(item);
assertEquals(item, dao.find(item.getId()));
}
}
I think your test are valid integration tests as stated above, but I would use Add to aid in the testing of of Find and vice verse..
At some level you have to allow yourself to place trust in your lowest level of integration to your external dependency. I realize there is a dependency to other methods in your tests, but I find that Add and Find methods are "low level" methods that are very easy to verify.
They essentially test each other as they are basically inverse methods.
Add can easily build preconditions for find
Find can verify that an add was successful.
I can't think of a scenario where a failure in either wouldn't be caught by your test
Your testAdd method has a problem: it depends on the assumption that ItemDao.list functions properly, and yet ItemDao is the Class that you're testing. Unit tests are meant to be independent, so a better approach is use plain JDBC -as #Amir said- to verify if the record was introduced in the database.
If you're using Spring, you can relay on AbstractTransactionalDataSourceSpringContextTests to access JDBCTemplate facilities and assure a rollback after the test was executed.
I use direct JDBC (using Spring's JdbcTemplate) to test the DAO methods. I mean I call the DAO methods (which are Hibernate base), and then confirm them using JDBC direct SQL calls.
The smallest unit under test for class-based unit testing is a class.
To see why, consider that you could, in theory, test each method of the class in isolation from all other methods by bypassing, stubbing or mocking them. Some tools may not support that; this is theory not practice, assume they do.
Even so, doing things that way would be a bad idea. The specification of an individual function by itself will vary between vaguely meaningless and verbosely incomprehensible. Only in the pattern of interaction between different functions will there exist a specification simpler than the code that you can profitably use to test it.
If you add an item and the number of items reported increases, things are working. If there is some way things could not be working, but nevertheless all the patterns of interaction hold, then you are missing some needed test.
Assume the following setup:
interface Entity {}
interface Context {
Result add(Entity entity);
}
interface Result {
Context newContext();
SpecificResult specificResult();
}
class Runner {
SpecificResult actOn(Entity entity, Context context) {
return context.add(entity).specificResult();
}
}
I want to see that the actOn method simply adds the entity to the context and returns the specificResult. The way I'm testing this right now is the following (using Mockito)
#Test
public void testActOn() {
Entity entity = mock(Entity.class);
Context context = mock(Context.class);
Result result = mock(Result.class);
SpecificResult specificResult = mock(SpecificResult.class);
when(context.add(entity)).thenReturn(result);
when(result.specificResult()).thenReturn(specificResult);
Assert.assertTrue(new Runner().actOn(entity,context) == specificResult);
}
However this seems horribly white box, with mocks returning mocks. What am I doing wrong, and does anybody have a good "best practices" text they can point me to?
Since people requested more context, the original problem is an abstraction of a DFS, in which the Context collects the graph elements and calculates results, which are collated and returned. The actOn is actually the action at the leaves.
It depends of what and how much you want your code to be tested. As you mentionned the tdd tag, I suppose you wrote your test contracts before any actual production code.
So in your contract what do you want to test on the actOn method:
That it returns a SpecificResult given both a Context and an Entity
That add(), specificResult() interactions happen on respectively the Context and the Entity
That the SpecificResult is the same instance returned by the Result
etc.
Depending on what you want to be tested you will write the corresponding tests. You might want to consider relaxing your testing approach if this section of code is not critical. And the opposite if this section can trigger the end of the world as we know it.
Generally speaking whitebox tests are brittle, usually verbose and not expressive, and difficult to refactor. But they are well suited for critical sections that are not supposed to change a lot and by neophytes.
In your case having a mock that returns a mock does look like a whitebox test. But then again if you want to ensure this behavior in the production code this is ok.
Mockito can help you with deep stubs.
Context context = mock(Context.class, RETURNS_DEEP_STUBS);
given(context.add(any(Entity.class)).specificResult()).willReturn(someSpecificResult);
But don't get used to it as it is usually considered bad practice and a test smell.
Other remarks :
Your test method name is not precise enough testActOn does tell the reader what behavior your are testing. Usually tdd practitioners replace the name of the method by a contract sentence like returns_a_SpecificResult_given_both_a_Context_and_an_Entity which is clearly more readable and give the practitioner the scope of what is being tested.
You are creating mock instances in the test with Mockito.mock() syntax, if you have several tests like that I would recommend you to use a MockitoJUnitRunner with the #Mock annotations, this will unclutter a bit your code, and allow the reader to better see what's going on in this particular test.
Use the BDD (Behavior Driven Dev) or the AAA (Arrange Act Assert) approach.
For example:
#Test public void invoke_add_then_specificResult_on_call_actOn() {
// given
... prepare the stubs, the object values here
// when
... call your production code
// then
... assertions and verifications there
}
All in all, as Eric Evans told me Context is king, you shall take decisions with this context in mind. But you really should stick to best practice as much as possible.
There's many reading on test here and there, Martin Fowler has very good articles on this matter, James Carr compiled a list of test anti-patterns, there's also many reading on using well the mocks (for example the don't mock types you don't own mojo), Nat Pryce is the co-author of Growing Object Oriented Software Guided by Tests which is in my opinion a must read, plus you have google ;)
Consider using fakes instead of mocks. It's not really clear what the classes in question are meant to to, but if you can build a simple in-memory (not thread-safe, not persistent etc) implementation of both interfaces, you can use that for flexible testing without the brittleness that sometimes comes from mocking.
I like to use names beginning mock for all my mock objects. Also, I would replace
when(result.specificResult()).thenReturn(specificResult);
Assert.assertTrue(new Runner().actOn(entity,context) == specificResult);
with
Runner toTest = new Runner();
toTest.actOn( mockEntity, mockContext );
verify( mockResult ).specificResult();
because all you're trying to assert is that specificResult() gets run on the right mock object. Whereas your original assert doesn't make it quite so clear what is being asserted. So you don't actually need a mock for SpecificResult. That cuts you down to just one when call, which seems to me to be about right for this kind of test.
But yes, this does seem frightfully white box. Is Runner a public class, or some hidden implementation detail of a higher level process? If it's the latter, then you probably want to write tests around the behaviour at the higher level; rather than probing implementation details.
Not knowing much about the context of the code, I would suggest that Context and Result are likely simple data objects with very little behavior. You could use a Fake as suggested in another answer or, if you have access to the implementations of those interfaces and construction is simple, I'd just use the real objects in lieu of Fakes or Mocks.
Although the context would provide more information, I don't see any problems with your testing methodology myself. The whole point of mock objects is to verify calling behavior without having to instantiate the implementations. Creating stub objects or using actual implementing classes just seems unnecessary to me.
However this seems horribly white box, with mocks returning mocks.
This may be more about the class design than the testing. If that is the way the Runner class works with the external interfaces then I don't see any problem with having the test simulate that behavior.
First off, since nobody's mentioned it, Mockito supports chaining so you can just do:
when(context.add(entity).specificResult()).thenReturn(specificResult);
(and see Brice's comment for how to do enable this; sorry I missed it out!)
Secondly, it comes with a warning saying "Don't do this except for legacy code." You're right about the mock-returning-mock being a bit strange. It's OK to do white-box mocking generally because you're really saying, "My class ought to collaborate with a helper like <this>", but in this case it's collaborating across two different classes, coupling them together.
It's not clear why the Runner needs to get the SpecificResult, as opposed to whatever other result comes out of context.add(entity), so I'm going to make a guess: the Result contains a result with some messages or other information and you just want to know whether it's a success or failure.
That's like me saying, "Don't tell me all about my shopping order, just tell me that I made it successfully!" The Runner shouldn't know that you only want that specific result; it should just return everything that came out, the same way that Amazon shows you your total, postage and all the things you bought, even if you've shopped there lots and are perfectly aware of what you're getting.
If some classes regularly use your Runner just to get a specific result while others require more feedback then I'd make two methods to do it, maybe called something like add and addWithFeedback, the same way that Amazon let you do one-click shopping by a different route.
However, be pragmatic. If it's readable the way you've done it and everyone understands it, use Mockito to chain them and call it a day. You can change it later if you have need.
How do you manage dummy data used for tests? Keep them with their respective entities? In a separate test project? Load them with a Serializer from external resources? Or just recreate them wherever needed?
We have an application stack with several modules depending on another with each containing entities. Each module has its own tests and needs dummy data to run with.
Now a module that has a lot of dependencies will need a lot of dummy data from the other modules. Those however do not publish their dummy objects because they are part of the test resources so all modules have to setup all dummy objects they need again and again.
Also: most fields in our entities are not nullable so even running transactions against the object layer requires them to contain some value, most of the time with further limitations like uniqueness, length, etc.
Is there a best practice way out of this or are all solutions compromises?
More Detail
Our stack looks something like this:
One Module:
src/main/java --> gets jared (.../entities/*.java contains the entities)
src/main/resources --> gets jared
src/test/java --> contains dummy object setup, will NOT get jared
src/test/resources --> not jared
We use Maven to handle dependencies.
module example:
Module A has some dummy objects
Module B needs its own objects AND the same as Module A
Option a)
A Test module T can hold all dummy objects and provide them in a test scope (so the loaded dependencies don't get jared) to all tests in all Modules. Will that work? Meaning: If I load T in A and run install on A will it NOT contain references introduced by T especially not B? Then however A will know about B's datamodel.
Option b)
Module A provides the dummy objects somewhere in src/main/java../entities/dummy allowing B to get them while A does not know about B's dummy data
Option c)
Every module contains external resources which are serialized dummy objects. They can be deserialized by the test environment that needs them because it has the dependency to the module to which they belong. This will require every module to create and serialize its dummy objects though and how would one do that? If with another unit test it introduces dependencies between unit tests which should never happen or with a script it'll be hard to debug and not flexible.
Option d)
Use a mock framework and assign the required fields manually for each test as needed. The problem here is that most fields in our entities are not nullable and thus will require setters or constructors to be called which would end us up at the start again.
What we don't want
We don't want to set up a static database with static data as the required objects' structure will constantly change. A lot right now, a little later. So we want hibernate to set up all tables and columns and fill those with data at unit testing time. Also a static data base would introduce a lot of potential errors and test interdependencies.
Are my thoughts going in the right direction? What's the best practice to deal with tests that require a lot of data? We'll have several interdependent modules that will require objects filled with some kind of data from several other modules.
EDIT
Some more info on how we're doing it right now in response to the second answer:
So for simplicity, we have three modules: Person, Product, Order.
Person will test some manager methods using a MockPerson object:
(in person/src/test/java:)
public class MockPerson {
public Person mockPerson(parameters...) {
return mockedPerson;
}
}
public class TestPerson() {
#Inject
private MockPerson mockPerson;
public testCreate() {
Person person = mockPerson.mockPerson(...);
// Asserts...
}
}
The MockPerson class will not be packaged.
The same applies for the Product Tests:
(in product/src/test/java:)
public class MockProduct() { ... }
public class TestProduct {
#Inject
private MockProduct mockProduct;
// ...
}
MockProduct is needed but will not be packaged.
Now the Order Tests will require MockPerson and MockProduct, so now we currently need to create both as well as MockOrder to test Order.
(in order/src/test/java:)
These are duplicates and will need to be changed every time Person or Product changes
public class MockProduct() { ... }
public class MockPerson() { ... }
This is the only class that should be here:
public class MockOrder() { ... }
public class TestOrder() {
#Inject
private order.MockPerson mockPerson;
#Inject
private order.MockProduct mockProduct;
#Inject
private order.MockOrder mockOrder;
public testCreate() {
Order order = mockOrder.mockOrder(mockPerson.mockPerson(), mockProduct.mockProduct());
// Asserts...
}
}
The problem is, that now we have to update person.MockPerson and order.MockPerson whenever Person is changed.
Isn't it better to just publish the Mocks with the jar so that every other test that has the dependency anyway can just call Mock.mock and get a nicely setup object? Or is this the dark side - the easy way?
This may or may not apply - I'm curious to see an example of your dummy objects and the setup code related. (To get a better idea of whether it applies to your situation.) But what I've done in the past is not even introduce this kind of code into the tests at all. As you describe, it's hard to produce, debug, and especially package and maintain.
What I've usaully done (and AFAIKT in Java this is the best practice) is try to use the Test Data Builder pattern, as described by Nat Pryce in his Test Data Builders post.
If you think this is somewhat relevant, check these out:
Does a framework like Factory Girl exist for Java?
make-it-easy, Nat's framework that implements this pattern.
Well, I read carefully all evaluations so far, and it is very good question. I see following approaches to the problem:
Set up (static) test data base;
Each test has it's own set up data that creates (dynamic) test data prior to running unit tests;
Use dummy or mock object. All modules know all dummy objects, this way there is no duplicates;
Reduce the scope of the unit test;
First option is pretty straight forward and has many drawbacks, somebody has to reproduce it's once in while, when unit tests "mess it up", if there are changes in the data-module, somebody has to introduce corresponding changes to the test data, a lot of maintenance overhead. Not to say that generation of this data on the first hand maybe tricky. See aslo second option.
Second option, you write your test code that prior to the testing invokes some of your "core" business methods that creates your entity. Ideally, your test code should be independent from the production code, but in this case, you will end up with duplicate code, that you should support twice. Sometimes, it is good to split your production business method in order to have entry point for your unit test (I makes such methods private and use Reflection to invoke them, also some remark on the method is needed, refactoring is now a bit tricky). The main drawback that if you must change your "core" business methods it suddenly effects all of your unit test and you can't test. So, developers should be aware of it and not make partials commits to the "core" business methods, unless they works. Also, with any change in this area, you should keep in your mind "how it will affect my unit test". Sometimes also, it is impossible to reproduce all the required data dynamically (usually, it is because of the third-parties API, for example, you call another application with it's own DB from which you required to use some keys. This keys (with the associated data) is created manually through third-party application. In such a case, this data and only this data, should be created statically. For example, your created 10000 keys starting from 300000.
Third option should be good. Options a) and d) sounds for me pretty good. For your dummy object you can use the mock framework or you can not to use it. Mock Framework is here only to help you. I don't see problem that all of your unit know all your entities.
Fourth option means that you redefine what is "unit" in your unit test. When you have couple of modules with interdependence than it can be difficult to test each module in isolation. This approach says, that what we originally tested was integration test and not unit test. So, we split our methods, extract small "units of works" that receives all it's interdependences to another modules as parameters. This parameters can be (hopefully) easily mocked up. The main drawback of this approach, that you don't test all of your code, but only so to say, the "focal points". You need to make integration test separately (usually by QA team).
I'm wondering if you couldn't solve your problem by changing your testing approach.
Unit Testing a module which depends on other modules and, because of that, on the test data of other modules is not a real unit test!
What if you would inject a mock for all of the dependencies of your module under test so you can test it in complete isolation. Then you don't need to setup a complete environment where each depending module has the data it needs, you only setup the data for the module your actually testing.
If you imagine a pyramid, then the base would be your unit tests, above that you have functional tests and at the top you have some scenario tests (or as Google calls them, small, medium and big tests).
You will have a huge amount of Unit Tests that can test every code path because the mocked dependencies are completely configurable. Then you can trust in your individual parts and the only thing that your functional and scenario tests will do is test if each module is wired correctly to other modules.
This means that your module test data is not shared by all your tests but only by a few that are grouped together.
The Builder Pattern as mentioned by cwash will definitely help in your functional tests.
We are using a .NET Builder that is configured to build a complete object tree and generate default values for each property so when we save this to the database all required data is present.